Brillouin adaptive self-supervised denoising: a denoising method trained without simulated dataset for Brillouin optical time domain analyzer

Brillouin optical time domain analyzer (BOTDA) is a distributed fiber optic sensor in which noise in the Brillouin gain spectra (BGS) significantly affects sensing range, spatial resolution, and Brillouin frequency shift (BFS) extraction accuracy. To maximize the utility of collected BGS and enhance BOTDA performance, we propose Brillouin Adaptive Self-supervised dEnoising (BASE), which we believe to be a novel denoising method based on self-supervised learning. BASE can directly utilize collected BGS data, generating sufficient training samples even from a limited number of noisy BGS images, to enable robust and efficient denoising. This approach circumvents the reliance on simulated noisy-clean BGS pairs required by supervised learning methods and overcomes the performance limitations of classical denoising methods. Experimental results demonstrate that BASE outperforms classical denoising methods such as anisotropic diffusion, 3D filtering, non-local means, and wavelet denoising. Furthermore, compared to the supervised learning method, BASE achieves a 9.5% reduction in root mean squared error (RMSE) for temperature extraction and a 25.4% alleviated spatial resolution (SR) deterioration. By innovating the learning process to eliminate reliance on simulated datasets and clean ground-truth BGS images, BASE effectively removes complex real-world noise, thereby improving sensor performance without requiring hardware modifications. Its ability to achieve high-fidelity noise reduction under varying noise levels and sampling intervals paves the way for more reliable and precise distributed fiber optic sensing, offering significant potential for practical applications.